부산대학교 도서관

Reinforcement learning is considered to be one of the main methods of general artificial intelligence, which can realize self-learning of machines through interaction with the environment. In this paper, a modified version of deep reinforcement learning algorithm based on the Actor-Critic framework is proposed. Unlike traditional updated methods, the algorithm proposed in this paper adopts a special on-policy method, which we called Accelerated Linear Approximation Method in Deep Actor-Critic Framework (ALA-AC). When the network is trained to a certain extent, the networks' parameters of some layers are frozen, and the remaining layers' parameters are trained for better strategy and faster training speed.